A proposed intervention threshold fuels debate over who controls AI risks.
WASHINGTON
Artificial intelligence developer Anthropic has presented US lawmakers with a proposed federal regulatory framework that would allow authorities to seek the suspension of advanced AI models under narrowly defined catastrophic-risk conditions. The draft has triggered criticism over whether its intervention thresholds would protect the public adequately or leave regulators unable to respond to serious threats before they escalate.
According to reporting by Politico and Reuters, the proposal would permit federal action when a model presents a foreseeable and imminent risk of causing more than 50 deaths or serious injuries, or property damage exceeding $1 billion.
These thresholds concern the potential suspension of an AI model, not a general requirement that regulators wait for casualties before conducting oversight. They would apply to specified threats involving biological weapons, cyberattacks or systems escaping their developers’ control.
The proposed system would require independent evaluators registered with the Department of Commerce to assess frontier AI developers and their models. Companies could finance these evaluations, provided compensation was not linked to the findings.
When evaluators identified a qualifying imminent threat, they would have 72 hours to notify the department. Federal authorities could then seek a suspension order from a judge in Washington.
That judicial requirement has become particularly controversial. Critics argue that the proposed structure could restrict the government’s ability to act against dangerous models, especially if other agencies were prevented from ordering an immediate suspension.
Nathan Calvin of the nonprofit organization Encode has questioned whether the draft would provide protections comparable to existing legislation. California state senator Scott Wiener has also challenged its implications for state-level AI safeguards.
A central concern involves federal preemption, the legal mechanism through which national legislation can supersede state laws. Critics fear that replacing stronger state protections with a narrower federal standard could reduce regulatory accountability.
Anthropic has acknowledged those objections. On October 18, federal affairs executive Anthony Cimino indicated that the company would revise provisions affecting state authority and reconsider the threshold involving potential casualties.
He nevertheless defended judicial review as a safeguard against excessive government intervention.
The debate reflects competing approaches to AI governance. One emphasizes preventing catastrophic events involving highly capable systems. Another argues that effective oversight must also address less dramatic but consequential harms, including failures affecting individuals and institutions.
Anthropic’s framework does not automatically resolve those differences. Nor has Congress enacted the proposal, which remains under discussion and revision.
The controversy also exposes questions about institutional independence when technology companies help design the rules governing their own products.
Technical expertise from developers can inform legislation, but public accountability depends on meaningful participation from researchers, regulators and affected communities.
For Washington, the challenge is determining when government intervention becomes justified without establishing thresholds so restrictive that credible warnings cannot produce timely action.
The unresolved issue is not simply whether 50 lives constitute an acceptable regulatory boundary. It is whether legislation can distinguish routine oversight from emergency powers while preserving the authority to prevent serious harm.
As advanced AI systems gain greater capabilities, that distinction may become central to the credibility of federal technology governance.
Beyond the Headlines, Accountability Matters.